prompt-eng-toolkit

Use when designing a new system/user prompt for an LLM task (refinement, classification, extraction, formatting, single-turn agent). Builds prompts using the v4 framework — persona+rationale → XML structure → outcome-first task body → injection defenses → output contract. Iterates against real provider APIs when an API key is available; falls back to theory-only design if not. Always validates against attack-fixture suite before finalizing.

Use when an existing prompt is too long, costs too much per call, has poor cache hit-rate, or models stopped adhering to its rules (often after a model version bump). Compresses tokens AND hardens adherence in one pass — these usually go together. Always produces a before/after token comparison table. Iterates against real provider APIs when an API key is available; falls back to theory-only review if not.